Motion Primitives

Breaking a complex motion into small, reusable, parameterizable building-block motions that get composed together when needed.

Motion primitives are small, reusable basic motions — ‘reach toward a spot,’ ‘pick up,’ ‘go forward then turn left.’ The idea is biologically inspired: continuous motion in humans and animals can be seen as a concatenation of a handful of basic segments. There are two main uses in robotics. One is in manipulation and imitation learning, representing a trajectory with a parameterized model learned from demonstrations, where execution just changes parameters like the target point or duration to transfer to a new situation — representative examples are dynamic movement primitives (DMP, generating a trajectory from a spring-damper system plus a learnable driving term) and probabilistic movement primitives (ProMP, learning a distribution over trajectories from multiple demonstrations). The other is in mobile-robot and self-driving-car planning: a set of short trajectories satisfying the vehicle's dynamics is precomputed, and a graph search such as A* strings them together online — this is called lattice planning. Compared to atomic skills, motion primitives operate at the trajectory level, while atomic skills more often refer to complete, semantically defined sub-tasks.

ExampleAutomated parking: a planner keeps a library of short trajectories a car can actually drive, such as ‘go straight a bit’ or ‘full steering lock while reversing 90°,’ and searches online with A* to string them into a route from the current position into the parking spot.

Also called

Movement Primitives

Related

Dynamic Movement Primitives · Probabilistic Movement Primitives · Skill Primitive · Imitation Learning · Kinodynamic Planning · A* Search

Sources

Movement Primitives in Robotics: A Comprehensive Survey (arXiv 2601.02379)
Spatio-Temporal Lattice Planning Using Optimal Motion Primitives (arXiv 2107.11467)